Abstract Introduction Remote monitoring (RM) data has been used to develop predictive algorithms that warn of decompensated heart failure before a heart failure episode. The same RM data could also be used to predict other adverse events in heart failure patients, such as ventricular arrhythmias. Machine learning (ML) models can be used to process complex multivariate time series RM data. Purpose The aim of this analysis was to generate a clinically useful model to predict defibrillator therapy (shock or anti-tachycardia pacing (ATP)). The output of the model is the likelihood of a shock in the next 24 hours. Methods RM data of all 292 implantable cardioverter defibrillators (ICDs) at our centre were collected (male 81%, 66±12 years). Electrograms of ICD therapies were manually verified. 63 time series variables were collated. Atrial lead-dependent RM parameters and ones with >30% missing data were excluded leaving 23 variables for analysis. After the exclusion of inappropriate therapies and patients with missing RM data, 218 patients were analysed. Of these, 41 (18.8%) had at least one ICD therapy. Three types of ML models were generated (1) random forest (RF), (2) extreme gradient boosted model (XGBoost), (3) one-dimensional convolutional neural network (1D-CNN) with three different intervals preceding ICD therapy: 30, 10, and 4 days. The model accuracy was measured utilising stratified grouped 5-fold cross-validation, with 20% of data used for testing in each fold. Results 4-day decision tree-based models (RF, XGBoost) demonstrated higher area under the curve (AUC) scores, compared with 10- and 30-day models. Percentage accuracy was similar across all three models and interval lengths. RF: The 4-day model had the highest AUC of 0.86 and an accuracy of 83.8%, compared to the 10-day (AUC 0.85, 83.0%) and 30-day (AUC 0.80, 82.2%) models. XGBoost: The 4-day model had an AUC of 0.85 and an accuracy of 83.4%, compared to the 10-day (AUC 0.84, 83.1%) and 30-day (AUC 0.81, 82.4%) models. 1D-CNN: The 4-day model had an AUC of 0.78 and an accuracy of 83.5%, compared to the 10-day (AUC 0.74, 83.1%) and 30-day (AUC 0.71, 82.2%) models. Episodes of prior ventricular tachycardia (VT) in the VT2 zone and prior ICD therapies contributed the most to model predictions, with their importance highest in the days immediately preceding ICD therapy (Figure 1). To better identify ‘first ICD therapy’, prior ICD therapies were excluded, and the models were re-run. AUC was lower at 0.70 with an accuracy of 72.9% in the highest-performing model (10-day XGBoost). The most important features were patient activity, RV pacing impedance, and mean ventricular heart rate (Figure 2). Conclusion We were able to predict ICD therapy with an AUC of 0.86. 4-day models were better than the 10-day and 30-day models. Model performance remained high (AUC 0.70), even after exclusion of prior ventricular arrhythmias from the predictors.1.Importance of top variables over time2.Top features without prior ICD therapy
Read full abstract
May 27, 2024
Ranu Rani Sethi + 5
Open Access